Google Reportedly Builds Gemini-Tuned 'Frozen v2' Chip for 2028

A chip designed to run one specific family of models, and only that family, is a stranger bet than the usual AI-hardware story. That is reportedly what Google is doing with a processor codenamed 'Frozen v2,' aimed squarely at its own Gemini models rather than at the general workload its TPUs currently carry, according to SiliconAngle's write-up of an original report in The Information.

The pitch is between six and 10 times better performance per watt than Google's current silicon. The efficiency is supposed to come from three places at once: fewer calculations required to run Gemini, less data shuttled back and forth between chip and memory, and a form of operator fusion that combines several calculations into a single computation. Google reportedly plans to fit enough memory on the chip to run Gemini fully on-chip, which is where the biggest chunk of the efficiency claim usually hides in these designs.

Why this matters if you don't build data centers: today's inference bill for a frontier model is dominated by moving weights and activations around silicon that was designed to be general. Freezing a specific architecture into the hardware trades flexibility for watts, and if the ratio really lands in the reported band, it changes what Google can charge, what it can serve at free-tier scale, and how aggressively it can undercut competitors renting Nvidia. Alphabet shares reportedly rose 1.5% on the news, which is the market's polite way of agreeing.

The honest caveat is that this is a single-sourced report about a chip that is not supposed to reach data centers until 2028. The reporting names nobody at Google, does not disclose a fab partner, and does not say whether Frozen v2 will supplement or eventually replace the TPUi inference part that sits alongside the training-focused TPU 8t. The performance figure should be read as an internal target, not a benchmark. And the whole strategy assumes Gemini's architecture stays stable enough by 2028 that etching it into silicon is a win rather than an anchor.

If that assumption holds, the interesting move is not the chip itself but the wager underneath it, that model design has matured enough to be worth freezing. That is a very different bet than the one Nvidia is making with a general-purpose stack, and it is the part worth watching.